Informal settlement expansion mapping for mortgage-portfolio and title-risk assessment
Unplanned urban settlements grow faster than any land registry can track, leaving mortgage lenders holding collateral whose legal status is genuinely uncertain. Very-high-resolution optical change detection and SAR coherence can map structure counts, roofing class and construction permanence at the parcel scale.
Sensors
- Maxar WorldView-3: Panchromatic resolution of 0.31 m and eight-band SWIR at 3.7 m. At this resolution individual roofing sheets, concrete slabs and mud-thatch panels are distinguishable, and new footprints as small as 15–20 m² can be detected. Revisit at any given point is roughly 1–4.5 days depending on latitude and tasking priority.
- Planet SuperDove: Eight spectral bands at 3–4 m resolution with a daily revisit cadence globally. Insufficient for individual roofing-material classification but well-suited to rapid change detection across large urban extents, flagging new construction clusters for targeted high-resolution follow-up.
- Sentinel-1 SAR (C-band): Interferometric coherence products derived from 6- or 12-day repeat passes at 10 m resolution. Coherence loss indicates surface disturbance; coherence persistence over multiple passes is a proxy for structural permanence. Free and open; archive extends to 2014.
- Capella Space X-band SAR: Spotlight mode delivers 0.35–0.5 m resolution SAR imagery. X-band backscatter is particularly sensitive to metallic roofing surfaces and building edge geometry, providing a structural-density signal independent of cloud cover or illumination. Tasking latency is typically under 24 hours.
Why the cadastre is the wrong place to look
In many emerging-market cities, 40 to 70 per cent of the urban population lives in settlements that predate or post-date formal land registration. A mortgage lender relying on cadastral polygons to bound its collateral is, in effect, using a photograph taken years or decades ago to describe a landscape that changes weekly. Incremental construction, plot subdivision and boundary encroachment are the norm, not the exception.
The risk is not merely legal abstraction. When a borrower defaults and a lender attempts to realise collateral, disputed boundaries and overlapping customary claims can stall enforcement for years. Portfolio-level exposure to this risk is almost impossible to quantify from registry data alone, because the registry simply does not record what has happened on the ground.
What a rooftop gives away about tenure security
Roofing material is an imperfect but useful proxy for a household's confidence in its own tenure. Corrugated iron sheets are cheap, portable and recoverable if eviction occurs. Concrete slabs require capital, time and a degree of certainty that the investment will not be demolished. Fired-brick walls under a concrete roof signal a different risk profile than timber frames under polythene sheeting, even if neither structure appears in any land register.
WorldView-3's eight SWIR bands (1195–2365 nm) provide spectral separability between metal, concrete, asbestos-cement and organic thatch that is not available from standard RGB or four-band imagery. Published studies using WorldView-2 and WorldView-3 data have demonstrated overall roofing-classification accuracies of 80–90 per cent across three to five material classes in sub-Saharan African and South Asian contexts. The honest caveat: separability collapses in the middle of the quality spectrum. Painted corrugated iron and low-grade fibre-cement tiles can produce overlapping spectral signatures, and a single mis-classification at the portfolio level can shift a risk band. Classification outputs should be treated as probabilistic scores, not binary categories.
SAR coherence: the permanence signal optical imagery cannot see
Optical imagery tells you what a structure looks like; SAR coherence tells you whether it has stayed put. In Sentinel-1 interferometric pairs, a coherence value close to 1.0 indicates that the surface scattering geometry has not changed between acquisitions. A freshly erected corrugated-iron structure on bare earth will show high coherence within weeks of completion, because metal is a stable radar reflector. A structure that is being modified, or one built from materials that shift with moisture, will show lower coherence over the same interval.
The practical application is a two-layer risk score. The optical layer classifies material quality; the SAR coherence layer assesses structural stability over time. A structure that scores mid-range on roofing material but shows high multi-temporal coherence across six or more Sentinel-1 passes is a different collateral proposition than one that scores similarly on optical classification but shows coherence instability. Capella X-band SAR adds a third dimension: at sub-metre resolution, double-bounce returns from building corners allow structure height estimation and wall-material inference, independent of cloud cover.
SAR is not without limits here. At Sentinel-1's 10 m resolution, individual structures in dense settlements are not resolved; the coherence signal is a neighbourhood-level aggregate. Capella X-band tasking resolves individual structures but costs money per acquisition and cannot provide the free, continuous time-series that Sentinel-1 delivers.
Measuring growth rates: the change-detection workflow
The core analytical product for a lender is a growth-rate map: how many new structures appeared in a defined settlement boundary over a defined period, and where are the highest-density expansion fronts. The standard approach pairs a baseline VHR image with a current image, applies a building-footprint extraction model (typically a convolutional neural network trained on labelled VHR data), and differences the two footprint layers.
Planet SuperDove's daily cadence is useful at the screening stage. A sudden increase in bare-earth pixels at the settlement fringe, visible in the red and near-infrared bands, flags active construction before a VHR acquisition is tasked. This two-stage approach keeps tasking costs proportionate to the risk signal.
Archive depth matters for lenders assessing historical exposure. Sentinel-1 data is publicly available from 2014 and WorldView archive imagery is commercially available in some areas from 2009 onwards. A lender can, in principle, reconstruct a settlement's growth trajectory over a decade and compare it against the origination date of loans in its portfolio.
Honest limits: what the data cannot settle
Satellite imagery cannot determine legal title. It can tell you that a structure exists, roughly when it appeared, what it is made of and whether it has remained stable. It cannot tell you who owns the underlying land, whether a customary claim is recognised under local law, or whether an eviction notice has been served. The output is evidence for a risk model, not a substitute for legal due diligence.
Cloud cover is a material constraint in tropical cities, which is precisely where informal settlement growth is fastest. Sentinel-1 and Capella SAR penetrate cloud, but optical classification of roofing material requires cloud-free acquisitions. In persistently cloudy environments, annual cloud-free composites may be the only reliable optical baseline, which limits the temporal resolution of the change signal to once or twice per year.
Building-footprint models trained in one city do not transfer without retraining to cities with different construction typologies. A model trained on West African compound housing performs poorly on South Asian incremental brick construction. Any deployment requires a locally validated training set.
From pixel to portfolio: packaging the output
The deliverable for a mortgage lender is not a satellite image. It is a structured risk layer that can be joined to a loan register by geographic coordinates. Each structure in the lender's collateral register gets a set of attributes: footprint area, roofing-material class with confidence score, multi-temporal coherence score, distance to the nearest formal cadastral boundary, and a settlement-growth-rate percentile for the surrounding 500-metre neighbourhood.
Satellize runs this workflow on open Sentinel-1 and Sentinel-2 data for baseline screening, with commercial VHR tasking added for high-priority areas where the screening signal is ambiguous. The same analytical architecture that underpins the Tonga crop-estimation programme, combining open-constellation time-series with targeted high-resolution tasking, applies directly to urban settlement monitoring.
For a lender with a geographically dispersed portfolio, the practical starting point is a screening pass across the full portfolio using Planet and Sentinel data to rank loan clusters by settlement-growth intensity. That ranking determines where WorldView-3 and Capella tasking budget is spent. The result is a proportionate use of commercial data rather than blanket coverage of every postcode.
Typical figures
| Highest optical resolution | 0.31 m panchromatic (WorldView-3); 3–4 m multispectral (Planet SuperDove) |
| SAR resolution | 10 m (Sentinel-1 IW mode); 0.35–0.5 m spotlight (Capella X-band) |
| Optical revisit | Daily (Planet SuperDove); 1–4.5 days per point (WorldView-3, tasked) |
| SAR revisit | 6 or 12 days (Sentinel-1); sub-24-hour tasking latency (Capella) |
| Spectral bands used | VIS, NIR, eight SWIR bands 1195–2365 nm (WorldView-3); C-band 5.4 GHz (Sentinel-1); X-band ~9.6 GHz (Capella) |
| Minimum detectable new structure | ~15–20 m² footprint at WorldView-3 resolution; neighbourhood-level aggregates only at Sentinel-1 10 m |
| Roofing-classification accuracy (published range) | 80–90 per cent overall for 3–5 material classes under cloud-free VHR conditions; lower in transitional material categories |
| Archive depth | Sentinel-1 from 2014 (open); WorldView commercial archive from ~2009 in some areas |
| Cloud penetration | Full (SAR); zero (optical classification requires cloud-free acquisition) |
| Delivery format | GeoPackage or GeoJSON footprint layers; GeoTIFF coherence rasters; CSV risk-attribute table joinable to loan register by coordinates |
Analytics Satellize can run
| Settlement growth-rate map | Multi-date building-footprint extraction via CNN on VHR optical imagery; differencing of footprint layers across baseline and current epochs | GeoJSON polygon layer with per-structure first-appearance date estimate and neighbourhood growth-rate percentile; updated on user-defined cadence |
| Roofing-material classification layer | Supervised spectral classification using WorldView-3 SWIR bands; training labels from field-verified or manually digitised samples; confidence score per polygon | GeoTIFF and vector layer with material class (metal, concrete, fibre-cement, organic, mixed) and per-pixel confidence; flagged ambiguous zones for field verification |
| Structural permanence score | Multi-temporal Sentinel-1 interferometric coherence stack (minimum six passes); mean and variance of coherence per structure cluster | Raster coherence time-series and per-neighbourhood permanence score (0–1 scale) in GeoPackage format |
| Portfolio collateral risk ranking | Spatial join of loan register coordinates to footprint, roofing-class, coherence and growth-rate layers; composite risk score weighted by lender-defined parameters | CSV table with one row per loan, including all satellite-derived attributes and a composite risk band (low / medium / high / data-gap); ready for credit-risk model ingestion |
| Settlement boundary change alert | Planet SuperDove daily change detection on bare-earth spectral index at settlement fringe; threshold-based alerting when expansion rate exceeds defined percentile | Email or API alert with bounding polygon of active expansion zone; triggers optional VHR tasking order |
| Historical growth trajectory report | Sentinel-1 and Sentinel-2 archive time-series analysis from 2014 to present; annual structure-count and density estimates | PDF and GIS report showing decadal growth curve for each settlement cluster in portfolio, with origination-date overlay |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.